Modelling injury severity of pedestrian crashes on urban road networks: a random-parameter spatial-joint logistic regression approach

Abstract

Pedestrian crashes on urban road networks are commonly modelled separately for intersections or road segments, potentially overlooking the network-level dependence between physically connected entities. This study developed a random-parameter spatial-joint logistic regression approach to jointly model pedestrian injury severityat intersections and on road segments within a unified framework. The approach captures cross-entity spatial dependence between intersections and road segments while accommodating unobserved heterogeneity. A case study used 716 pedestrian crashes, 204 intersections, and 366 road segments in Wan Chai, Hong Kong, China. Accounting for cross-entity spatial dependence improved both model goodness-of-fit and predictive performance. Ablation analysis confirmed additional gains from incorporating spatial correlation into the random-parameter framework. Older pedestrians and those with head injuries had higher fatal or severe injury risks in both settings. At intersections, unmarked crossings, newer vehicles, and congestion were negatively associated with severity; on road segments, pedestrian inattention and morning crashes were associated with greater severity.